Recruiters sit on hundreds of CVs they can't search. Keyword search misses the obvious — the candidate wrote "React", you searched "frontend", they never surface — and every AI tool that fixes that ships candidate data to a third-party API.

I built RecruitSecure AI to do the semantic search without the API call. The embedding model runs server-side through Transformers.js (quantized multilingual-e5-small, 384 dimensions, 100+ languages), so you can ask in Spanish and it finds CVs written in English — and nothing ever leaves the deployment.

Drop in PDFs or DOCX and it extracts name, email, skills and experience, indexes each CV as a vector, and answers plain-language queries like "frontend lead who has built a design system": ranked by relevance, each result showing the exact CV fragments behind the match, so the recruiter sees the reason and not just a score. Retrieval is hybrid — pgvector HNSW cosine fused with BM25 full-text through Reciprocal Rank Fusion, then re-ranked from the recruiter's own thumbs up/down.

Next.js 16, React 19, strict TypeScript, PostgreSQL + pgvector through Drizzle ORM, Auth.js, Stripe subscriptions, multi-tenant with AES-256-GCM encryption at rest and GDPR export and deletion. Built solo, end to end.

The video is the real application: ten synthetic CVs uploaded through the actual pipeline, extracted, embedded and searched. Nothing staged.